DIC speckle image preparation method based on convolutional neural network
The Speckle Style Transfer Image Generation Network (SSTIGNet) constructed through convolutional neural networks solves the problems of precise control and equipment dependence in DIC speckle preparation, and achieves efficient generation of high-quality speckle images, which is suitable for digital image correlation measurements in multiple fields.
Patent Information
- Application Number
- CN202510672986.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-23
- Publication Date
- 2025-09-05
AI Technical Summary
Existing DIC speckle preparation methods have problems such as difficulty in accurately controlling spot size and density, high equipment dependence and environmental interference, making it difficult to stably and efficiently prepare high-quality speckle patterns.
A convolutional neural network is used to construct a speckle style transfer image generation network (SSTIGNet). Through the style transfer module and the difference extraction module, high-quality DIC speckle images are generated. Technical means such as dual encoders, adaptive convolution and optical flow encoders are included to achieve precise control of speckle features and difference extraction.
It achieves efficient generation of high-quality speckle images for arbitrary image pairs, overcomes the limitations of traditional methods, improves the controllability and flexibility of speckle image preparation, and has strong adaptability, making it suitable for fields such as materials science, mechanical engineering, and biomedical imaging.
Smart Images
Figure CN120598769A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of digital image correlation (DIC) and deep learning, and neural style transfer (NST) and convolutional neural networks (CNN), and in particular to a method for preparing DIC speckle images based on a convolutional neural network. Background Art
[0002] Convolutional neural networks are a common and efficient basic network architecture in the field of deep learning. They achieve hierarchical extraction by extracting features through convolutional layers and reducing resolution through downsampling through pooling layers. They are particularly widely used in computer vision problems and possess powerful image processing capabilities. Neural style transfer is a key research area in computer vision. It can transfer the artistic style characteristics of one style image to another content image while preserving the overall content. It is an applied technology for image feature extraction and generation. Therefore, style transfer models built on convolutional neural networks have been widely used in the field of neural style transfer.
[0003] In recent years, digital image correlation (DIC) has emerged as a prominent nondestructive optical measurement technique, characterized by its non-contact nature, full-field measurement capability, high precision, and adaptability to diverse environments. It obtains the displacement vector of a pixel by matching its position before and after deformation, thereby generating a full-field displacement map. This has led to its widespread application in fields such as materials science, mechanical engineering, civil engineering, and biomedical imaging. Currently, the main methods for generating DIC speckle patterns include spraying, pasting, and optical projection, each with its own unique advantages and limitations. While spraying speckle patterns is simple and easy to implement, precise control of spot size and density presents challenges. Pasting methods can accurately control spot size, shape, and density, but their application is limited. Optical projection methods offer rapid speckle generation through surface projection, but require specialized equipment, are susceptible to interference from ambient light, and can produce decorrelation effects. Currently, the stable generation of high-quality speckle patterns and the resolution of these limitations remain pressing challenges.
[0004] Compared with artificial speckle images, computer-simulated speckle image generation enables precise control of speckle characteristics, including particle size, contrast, and sharpness, through parameter adjustment, offering significant advantages including enhanced controllability and improved flexibility. Summary of the Invention
[0005] To address the shortcomings of the aforementioned speckle preparation methods, the present invention introduces a convolutional neural network for DIC speckle production. A convolutional neural network-based DIC speckle image preparation method is proposed, and a novel neural network framework, called the Speckle Style Transfer Image Generation Network (SSTIGNet), is constructed. This method aims to overcome the limitations of traditional methods and stably and efficiently produce high-quality DIC speckle patterns.
[0006] To achieve the above objectives, the present invention proposes a DIC speckle image preparation method based on a convolutional neural network, comprising the following steps:
[0007] Step 1: Input the object's surface image before deformation, the surface image after deformation, and the speckle example image as the reference image, style image, and deformed image, respectively.
[0008] Step 2: Use the style transfer network to extract style features and prepare reference speckles.
[0009] Step 3: Use the difference extraction network to extract the difference information between the reference image and the deformed image to prepare the deformed speckle.
[0010] See Figure 2 Furthermore, step 2 extracts and processes features of the reference image and the style image, and then generates reference speckles through the style transfer network, specifically including:
[0011] Step 2.1: Use a dual encoder (content encoder and style feature encoder) to encode the reference image and style image respectively.
[0012] Step 2.2: Generate two sets of high-frequency and low-frequency content features and style features through the encoder.
[0013] The content encoder and style feature encoder are developed based on the enhanced MobileNet framework. By integrating Octave convolution (OctConv) into the depthwise separable convolution module, the network implements frequency-based feature map decomposition, effectively reducing computational redundancy while maintaining network simplicity and efficiency.
[0014] Specifically, the spatial reduction mechanism in the low-frequency components of octave convolution expands the receptive field, enabling the network to extract contextual information from a wider area, especially improving performance at high resolution. At the same time, a notable feature of octave convolution is its ability to maintain efficient information exchange between high-frequency and low-frequency components while performing frequency decomposition. In addition, the high-frequency components of octave convolution retain fine-grained details including textures and local patterns, while the low-frequency components maintain global structural information and tonal relationships. And octave convolution optimizes spatial efficiency by reducing the spatial resolution of low-frequency feature maps.
[0015] Step 2.3: Use the kernel prediction network to use two sets of high-frequency and low-frequency style feature descriptors to predict the style-aware convolution kernel and bias at each spatial resolution level.
[0016] Among them, style features are integrated into the generator through adaptive convolution (AdaConv), and the fusion of content and style features is achieved with the help of the kernel prediction network. The kernel prediction network analyzes high-frequency and low-frequency style features in a deeply separable manner to generate kernels and biases.
[0017] Step 2.4: Feature fusion is performed by combining the content features with their corresponding frequency-specific style-aware convolutional kernels and biases through the generator.
[0018] The generator consists of three layers, each of which contains an adaptive convolution module, a standard octave convolution module and an upsampling module.
[0019] Specifically, the octave convolution module is designed to learn style-independent kernels, which is crucial for reconstructing high-fidelity images.
[0020] Step 2.5: The style transfer network generates a reference speckle pattern containing the required speckle example image features by hierarchically integrating and synthesizing the output of high- and low-frequency components.
[0021] Specifically, through the sequential processing of three generator layers, content and style features are effectively fused to generate reference speckles.
[0022] Furthermore, step 3 extracts difference information between the reference image and the deformed image, and then generates deformed speckle through a difference extraction network, which specifically includes:
[0023] Step 3.1: The optical flow encoder extracts feature representation and context information from the input reference image and deformed image pair.
[0024] Among them, the differential feature encoder consists of two convolutional networks, a feature encoder and a context encoder. Each encoder contains two convolutional layers and three residual layers, and no weights are shared between them.
[0025] Specifically, the feature encoder processes the input image pair I1 and I2 to generate feature maps F1 and F2 with a resolution of 1 / 8 of the input image, which are then used for correlation calculations. The contextual features are combined with the relevant outputs aggregated by the cross-strip module and input into the update block to iteratively refine the differential features at a fixed resolution.
[0026] Step 3.2: Calculate the orthogonal feature correlations through the cross-strip module (CSC) and integrate them with the correlations of all pairs to initialize the difference features.
[0027] Among them, the CSC module extracts the orthogonal query matrix Q from the target image v and Q u , and extracting the orthogonal bond matrix from the reference image and The striping operation ensures the directional consistency of global information. The correlation between these matrices is calculated to generate the global visual similarity index of I1 and I2 in the vertical and horizontal directions.
[0028] Step 3.3, the Correlation Regression Initialization module (CRI) further improves this initialization process by optimizing the CSC output.
[0029] The CRI module effectively utilizes the orthogonal correlation C derived from the CSC framework. v and C u , given the orthogonal correlation volume C v and C u ,We first apply the softmax transform to activate each volume separately, and then perform element-wise multiplication with the respective original correlation volume to generate an orthogonal energy map.
[0030] In step 3.4, the update block synthesizes the final difference features using the contextual information of the reference image, the aggregated correlation volume (combining orthogonality and correlation of all pairs), and the initialized difference features generated by CRI.
[0031] Among them, our framework utilizes convolutional GRU as the update module to iteratively optimize the differential features.
[0032] Specifically, the update module processes three main inputs: (1) the context feature representation C1 extracted from I1 by the context encoder; (2) the context feature representation C2 extracted from I1 by the CSC module by connecting C and C v and C u Generated aggregate correlation volume (3) Initial differential features V0 generated by the CRI module.
[0033] Step 3.5: The difference extraction network generates a deformed speckle pattern containing the required speckle example image features by integrating the difference feature synthesis output.
[0034] This study proposes a convolutional neural network-based method for DIC speckle image preparation, involving a groundbreaking Speckle Style Transfer Image Generation Network (SSTIGNet) framework specifically designed to efficiently generate paired reference and deformed speckle images from any input image pair, guided by a single reference speckle example. The proposed SSTIGNet comprises two functionally distinct but complementary modules: a style transfer module and a difference extraction module. The style transfer module integrates the speckle features of the target speckle image into the reference image while preserving the essential texture properties of the reference image to generate a reference speckle image. Conversely, the difference extraction module quantitatively analyzes the deformation features between the input reference image and its deformed image, and then embeds these deformation features into the style-transferred reference image to generate the final deformed speckle image. The reference and deformed speckle pairs generated using this method are highly useful for digital image correlation measurements.
[0035] The advantage of the present invention is that this method overcomes the key limitations of traditional methods, especially the inefficiency of manual preparation and the challenge of difficult quality control, and can process any image pair to generate corresponding speckle images, representing an important progress in the field of speckle image preparation. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 This is a flow chart of the DIC speckle image preparation process based on a convolutional neural network according to the present invention.
[0037] Figure 2 This is a model structure diagram of the present invention.
[0038] Figure 3 This is the core framework diagram of the content and style feature encoder of the present invention.
[0039] Figure 4 Kernel prediction network and generator graph provided by the present invention.
[0040] Figure 5 This is a diagram of the CSC module and CRI module provided by the present invention.
[0041] Figure 6 This is a diagram of the update block module provided by the present invention.
[0042] Figure 7 Comparison images of reference speckles generated by our method and those generated by other style transfer models are shown.
[0043] Figure 8 Figure 2 is a comparison of the reference speckle and deformed speckle generated by our method. DETAILED DESCRIPTION
[0044] To make the principles, features and advantages of the present invention more clearly understood, the following describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. It should be emphasized that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0045] Figure 1 This is a flow chart of the DIC speckle image preparation process based on a convolutional neural network of the present invention, see Figure 1 The present invention proposes a DIC speckle image preparation method based on a convolutional neural network, comprising the following steps:
[0046] Step 1: Input the object's surface image before deformation, the surface image after deformation, and the speckle example image as the reference image, style image, and deformed image, respectively.
[0047] Step 2: Use the style transfer network to extract style features and prepare reference speckles.
[0048] Step 3: Use the difference extraction network to extract the difference information between the reference image and the deformed image to prepare the deformed speckle.
[0049] Figure 2 For the model structure diagram of the present invention, see Figure 2 Furthermore, step 2 extracts and processes features of the reference image and the style image, and then generates reference speckles through the style transfer network, specifically including:
[0050] Step 2.1: Use a dual encoder (content encoder and style feature encoder) to encode the reference image and style image respectively.
[0051] Step 2.2: Generate two sets of high-frequency and low-frequency content features and style features through the encoder.
[0052] Figure 3 The core framework diagram of the content and style feature encoder of the present invention is shown in FIG. Figure 3 The content encoder and style feature encoder are developed based on the enhanced MobileNet framework. By integrating octave convolution into the depthwise separable convolution module, the network achieves frequency-based feature map decomposition, effectively reducing computational redundancy while maintaining the simplicity and efficiency of the network.
[0053] Specifically, the spatial reduction mechanism in the low-frequency components of octave convolution expands the receptive field, enabling the network to extract contextual information from a wider area, especially improving performance at high resolution. At the same time, a notable feature of octave convolution is its ability to maintain efficient information exchange between high-frequency and low-frequency components while performing frequency decomposition. In addition, the high-frequency components of octave convolution retain fine-grained details including textures and local patterns, while the low-frequency components maintain global structural information and tonal relationships. And octave convolution optimizes spatial efficiency by reducing the spatial resolution of low-frequency feature maps.
[0054] See Figure 3 , specifically, for a given input X={X H , X L} and output Y={Y H , Y L}, the forward propagation transformation can be expressed as:
[0055] Y H =f(X H ;W H→H )+f(upsample(X L ,2);W L→H ) (1)
[0056] Y L =f(X L ;W L→L )+f(pool(X H ,2);W H→L ) (2)
[0057] where f(X;W) denotes a convolution operation parameterized by W, pool(X,2) corresponds to average pooling using a 2×2 kernel, and upsample(X,2) denotes an upsampling operation with a scaling factor of 2 achieved via nearest neighbor interpolation.
[0058] Step 2.3: Use the kernel prediction network to use two sets of high-frequency and low-frequency style feature descriptors to predict the style-aware convolution kernel and bias at each spatial resolution level.
[0059] Figure 4 For the kernel prediction network and generator diagram provided by the present invention, see Figure 4 Among them, the style features are integrated into the generator through adaptive convolution, and the fusion of content and style features is achieved with the help of the kernel prediction network. The kernel prediction network analyzes high-frequency and low-frequency style features in a deep separable way to generate kernels and biases.
[0060] Step 2.4: Feature fusion is performed by combining the content features with their corresponding frequency-specific style-aware convolutional kernels and biases through the generator.
[0061] See Figure 4 , where the generator consists of three layers, each of which contains an adaptive convolution module, a standard octave convolution module and an upsampling module.
[0062] Specifically, the octave convolution module is designed to learn style-independent kernels, which is crucial for reconstructing high-fidelity images.
[0063] Step 2.5: The style transfer network generates a reference speckle pattern containing the required speckle example image features by hierarchically integrating and synthesizing the output of high- and low-frequency components.
[0064] Specifically, through the sequential processing of three generator layers, content and style features are effectively fused to generate reference speckles.
[0065] We use the pre-trained VGG-19 network to calculate the content-aware loss L c and style-aware loss L s In addition, the style-aware loss L is redefined based on the Exact Feature Distribution Matching (EFDM) algorithm. s , assuming I c Represents a stylized image, I c represents the content image. The perceptual loss is calculated as follows:
[0066] L c =||f3(I cs )-f3(I c )||2 (3)
[0067]
[0068] Among them, f n Represents the nth layer of the VGG-19 model. Content-aware loss L c It is calculated in the conv3_1 layer of VGG-19, and the style perception loss L s It is calculated between conv1_1, conv2_1, conv3_1 and conv4_1 layers. It is worth noting that the VGG-19 model is only used during training.
[0069] When a CNN-based style transfer network repeatedly stylizes the same pair of content and style images, it inevitably suffers from content leakage. This content leakage can lead to the loss of details in the stylized image, causing local blurring. This is particularly disadvantageous in DIC techniques, which require extremely high-quality spot images. To address this issue, we introduced the Structural Similarity Index (SSIM) function and L1 loss during the neural network training process:
[0070]
[0071] Where x and y represent I cs and I s The local area or pixel block in μ x and μ y represents the average value of x and y, representing brightness; and and Then they represent their variance and covariance σ xy Capturing the structural similarity between x and y, constants C1 and C2 are included to avoid division by zero. si and I csi Respectively represent I s and I cs The value of the i-th pixel in , N represents the total number of pixels in the image.
[0072] Since the range of SSIM is -1 to 1, a value close to 1 indicates a higher similarity between two images, and a smaller L1 loss value corresponds to a higher image similarity. To optimize the model, we define the following loss function:
[0073] L SSIM =1-SSIM(x,y) (7)
[0074] L L1 =L1(I s ,I cs ) (8)
[0075] By integrating all the above losses, the overall loss function is defined as follows:
[0076] L total =λ c L c +λ s L s +λ SSIM L SSIM +λ L1 L L1 (9)
[0077] Among them, λ c ,λ s ,λ SSIM and λ L1 Represent the weight of each loss term respectively. In this study, we set these hyperparameter settings to 1, 10, 100, 100 respectively to alleviate the impact of content leakage on stylized images.
[0078] Furthermore, we use MS-COCO and WikiArt as content and style datasets, respectively, and randomly select 60,000 images from each dataset to construct the training set. During training, the images are resized to 512 pixels and then randomly cropped to 128×128 pixels for data augmentation. The model is optimized using the Adam optimizer with a learning rate of 0.0001, a batch size of 16, and 160,000 iterations.
[0079] See Figure 2 ,Furthermore, step 3 extracts difference information between the reference image and the deformed image, and then generates the deformed speckle through the difference extraction network, specifically including:
[0080] Step 3.1: The optical flow encoder extracts feature representation and context information from the input reference image and deformed image pair.
[0081] Among them, the differential feature encoder consists of two convolutional networks, a feature encoder and a context encoder. Each encoder contains two convolutional layers and three residual layers, and no weights are shared between them.
[0082] Specifically, the feature encoder processes the input image pair I1 and I2 to generate feature maps F1 and F2 with a resolution of 1 / 8 of the input image, which are then used for correlation calculations. The contextual features are combined with the relevant outputs aggregated by the cross-strip module and input into the update block to iteratively refine the differential features at a fixed resolution.
[0083] Step 3.2: Calculate the orthogonal feature correlations through the cross-strip module (CSC) and integrate them with the correlations of all pairs to initialize the difference features.
[0084] Figure 5 For the CSC module and CRI module diagram provided by the present invention, refer to Figure 5 , where the CSC module extracts the orthogonal query matrix Q from the target image v and Q u , and extracting the orthogonal bond matrix from the reference image and The striping operation ensures the directional consistency of global information. The correlation between these matrices is calculated to generate the global visual similarity index of I1 and I2 in the vertical and horizontal directions.
[0085] Specifically, the feature encoder processes the input reference image I1 and the deformed image I2 to generate feature maps F1 and F2, and applies two 1×1 convolutional layers to F1 to generate horizontal and vertical query matrices Q respectively. v and Q u At the same time, F2 is processed by two additional 1×1 convolutional layers to obtain the key matrix Kv and K u , these matrices are further refined by vertical and horizontal striping operations to finally obtain the orthogonal global bond matrix and As shown below:
[0086]
[0087] Specifically, the orthogonal query matrix is transposed to convert its dimensions from C×H×W to H×W×C. The transposed matrix is dot-producted with the orthogonal key value to generate two different correlation volumes C v and C u , these correlation volumes encode non-local visual information similarity, thereby capturing vertical and horizontal spatial relationships in the feature space.
[0088] Step 3.3, the Correlation Regression Initialization module (CRI) further improves this initialization process by optimizing the CSC output.
[0089] See Figure 5 , the CRI module effectively utilizes the orthogonal correlation C derived from the CSC framework v and C u , given the orthogonal correlation volume C v and C u ,We first apply the softmax transform to activate each volume separately, and then perform element-wise multiplication with the respective original correlation volume to generate an orthogonal energy map.
[0090] Specifically, this process ultimately produces vertical and horizontal differential feature initializations, denoted as v0 and u0, respectively, as shown below:
[0091]
[0092] In step 3.4, the update block synthesizes the final difference features using the contextual information of the reference image, the aggregated correlation volume (combining orthogonality and correlation of all pairs), and the initialized difference features generated by CRI.
[0093] Figure 6 For the update block module diagram provided by the present invention, refer to Figure 6 , where our framework utilizes convolutional GRU as the update module to iteratively optimize the differential features.
[0094] Specifically, the update module processes three main inputs: (1) the context feature representation C1 extracted from I1 by the context encoder; (2) the context feature representation C2 extracted from I1 by the CSC module by connecting C and C v and C u Generated aggregate correlation volume (3) Initial differential features V0 generated by the CRI module.
[0095] Specifically, C is calculated by the tensor dot product operation between the feature tensors F1 and F2, and the formula is as follows:
[0096] C(x1,y1,x2,y2)=F1(x1,y1) T ·F2(x2,y2) (14)
[0097] The update block iteratively estimates a differential feature sequence {V1, V2, ..., V m}, where m=12. During training, at each iteration, the incremental flow field ΔV is calculated and added to the current estimate so that V i+1 =V i +ΔV, after each update, V i Upsampled to match the input resolution.
[0098] Step 3.5: The difference extraction network generates a deformed speckle pattern containing the required speckle example image features by integrating the difference feature synthesis output.
[0099] Furthermore, we use the weighted sequential loss function to calculate the predicted optical flow V i and the true value V t The error L between them, the initial estimate V0 is provided by the CRI module, and the subsequent prediction values {V1, V2, ..., V m Supervision in the following manner:
[0100]
[0101] Among them, γ=0.8 and m=12.
[0102] Furthermore, we used a model trained with CSFlow to analyze the difference features between the reference image and the deformed image. The model was initially trained on the FlyingChairs dataset for 150,000 iterations with a batch size of 10 and a learning rate of 4e -4 , the cropped image size is 368 × 496. Subsequently, the model is trained on the FlyingThings dataset for 150,000 iterations with a batch size of 6 and a learning rate of 1.25e -4 , the cropped image size is 400 × 720. Finally, the model is fine-tuned on the KITTI dataset.
[0103] Figure 7 A comparison of the reference speckles generated by our method and those generated by other style transfer models is shown in Figure 2. Figure 7 , Figure 7Comparison images of reference speckle generated by our method and other style transfer models are shown. Clearly, for three different sets of reference and style images, the reference speckle generated by our model exhibits a relatively uniform distribution and high-contrast speckle pattern, exhibiting higher quality than other models, despite lacking significant white spots, edge loss, or blurring.
[0104] Figure 8 For a comparison of the reference and deformed speckles generated by our method, see Figure 8 , Figure 8 The reference and deformed speckle pairs generated using our method are shown. Results from four different input image groups, Groups 1 to 4, show excellent correspondence between the reference and deformed speckle pairs generated by our method. The deformed speckle and reference speckle patterns exhibit only image deformation due to the displacement field, and exhibit good correlation in terms of content, style, and detail. This significantly aids in DIC measurement using these pairs.
[0105] This study proposes a convolutional neural network-based method for DIC speckle image preparation, involving a groundbreaking Speckle Style Transfer Image Generation Network (SSTIGNet) framework specifically designed to efficiently generate paired reference and deformed speckle images from any input image pair, guided by a single reference speckle example. The proposed SSTIGNet comprises two functionally distinct but complementary modules: a style transfer module and a difference extraction module. The style transfer module integrates the speckle features of the target speckle image into the reference image while preserving the essential texture properties of the reference image to generate a reference speckle image. Conversely, the difference extraction module quantitatively analyzes the deformation features between the input reference image and its deformed image, and then embeds these deformation features into the style-transferred reference image to generate the final deformed speckle image. The reference and deformed speckle pairs generated using this method are highly useful for digital image correlation measurements.
[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the principles and principles of the present invention that can be easily thought of by any technician familiar with this technical field should be included in the scope of protection of the present invention.
Claims
1. A method for preparing DIC speckle images based on convolutional neural networks, characterized in that: The steps include: Step 1: Input the surface image of the object before deformation, the surface image after deformation, and the speckle example image as the reference image, style image, and deformed image respectively; Step 2: Use the style transfer network to extract style features and prepare reference speckles; Step 3: Use the difference extraction network to extract the difference information between the reference image and the deformed image to prepare the deformed speckle.
2. The method for preparing a DIC speckle image based on a convolutional neural network as claimed in claim 1, wherein: Step 2 includes the following steps: Step 2.1: Use a dual encoder (content encoder and style feature encoder) to encode the reference image and style image respectively; Step 2.2: Generate two sets of high-frequency and low-frequency content features and style features through the encoder; Step 2.3: Use the kernel prediction network to predict the style-aware convolution kernel and bias for each spatial resolution level using the high-frequency and low-frequency style feature descriptors. Step 2.4: Perform feature fusion by combining content features with their corresponding frequency-specific style-aware convolutional kernels and biases through the generator. Step 2.5: The style transfer network generates a reference speckle pattern containing the required speckle example image features by hierarchically integrating and synthesizing the output of high- and low-frequency components.
3. The method for preparing DIC speckle images based on a convolutional neural network as claimed in claim 1, wherein: Step 3 includes the following steps: Step 3.1, the optical flow encoder extracts feature representation and context information from the input reference image and deformed image pair; Step 3.2: Calculate the orthogonal feature correlations through the cross-strip module (CSC) and integrate them with the correlations of all pairs to initialize the difference features; Step 3.3, the Correlation Regression Initialization module (CRI) further improves this initialization process by optimizing the CSC output; Step 3.4: The update block uses the contextual information of the reference image, the aggregated correlation volume (combining orthogonality and correlation of all pairs), and the initialized difference features generated by CRI to synthesize the final difference features; Step 3.5: The difference extraction network generates a deformed speckle pattern containing the required speckle example image features by integrating the difference feature synthesis output.
4. The method for preparing DIC speckle images based on a convolutional neural network as claimed in claim 1, wherein: The input object surface image before deformation, surface image after deformation, and speckle example image are used as reference image, style image, and deformed image respectively.
5. The method for preparing DIC speckle images based on a convolutional neural network as claimed in claim 1, wherein: The model has two major modules, the style transfer module and the difference extraction module. The style transfer module is used to prepare reference speckles, and the difference extraction module is used to prepare deformed speckles.